lustre/include/linux/lustre_dlm.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * (visit-tags-table FILE)
   3  * vim:expandtab:shiftwidth=8:tabstop=8:
   4  */
   5
   6 #ifndef _LUSTRE_DLM_H__
   7 #define _LUSTRE_DLM_H__
   8
   9 #ifdef __KERNEL__
  10 # include <linux/proc_fs.h>
  11 #endif
  12
  13 #include <linux/lustre_lib.h>
  14 #include <linux/lustre_net.h>
  15 #include <linux/lustre_import.h>
  16 #include <linux/lustre_handles.h>
  17 #include <linux/lustre_export.h> /* for obd_export, for LDLM_DEBUG */
  18
  19 struct obd_ops;
  20 struct obd_device;
  21
  22 #define OBD_LDLM_DEVICENAME  "ldlm"
  23
  24 #define LDLM_DEFAULT_LRU_SIZE 100
  25
  26 typedef enum {
  27         ELDLM_OK = 0,
  28
  29         ELDLM_LOCK_CHANGED = 300,
  30         ELDLM_LOCK_ABORTED = 301,
  31         ELDLM_LOCK_REPLACED = 302,
  32
  33         ELDLM_NAMESPACE_EXISTS = 400,
  34         ELDLM_BAD_NAMESPACE    = 401
  35 } ldlm_error_t;
  36
  37 #define LDLM_NAMESPACE_SERVER 0
  38 #define LDLM_NAMESPACE_CLIENT 1
  39
  40 #define LDLM_FL_LOCK_CHANGED   0x000001 /* extent, mode, or resource changed */
  41
  42 /* If the server returns one of these flags, then the lock was put on that list.
  43  * If the client sends one of these flags (during recovery ONLY!), it wants the
  44  * lock added to the specified list, no questions asked. -p */
  45 #define LDLM_FL_BLOCK_GRANTED  0x000002
  46 #define LDLM_FL_BLOCK_CONV     0x000004
  47 #define LDLM_FL_BLOCK_WAIT     0x000008
  48
  49 #define LDLM_FL_CBPENDING      0x000010 /* this lock is being destroyed */
  50 #define LDLM_FL_AST_SENT       0x000020 /* blocking or cancel packet was sent */
  51 #define LDLM_FL_WAIT_NOREPROC  0x000040 /* not a real flag, not saved in lock */
  52 #define LDLM_FL_CANCEL         0x000080 /* cancellation callback already run */
  53
  54 /* Lock is being replayed.  This could probably be implied by the fact that one
  55  * of BLOCK_{GRANTED,CONV,WAIT} is set, but that is pretty dangerous. */
  56 #define LDLM_FL_REPLAY         0x000100
  57
  58 #define LDLM_FL_INTENT_ONLY    0x000200 /* don't grant lock, just do intent */
  59 #define LDLM_FL_LOCAL_ONLY     0x000400 /* see ldlm_cli_cancel_unused */
  60
  61 /* don't run the cancel callback under ldlm_cli_cancel_unused */
  62 #define LDLM_FL_NO_CALLBACK    0x000800
  63
  64 #define LDLM_FL_HAS_INTENT     0x001000 /* lock request has intent */
  65 #define LDLM_FL_CANCELING      0x002000 /* lock cancel has already been sent */
  66 #define LDLM_FL_LOCAL          0x004000 /* local lock (ie, no srv/cli split) */
  67 #define LDLM_FL_WARN           0x008000 /* see ldlm_cli_cancel_unused */
  68 #define LDLM_FL_DISCARD_DATA   0x010000 /* discard (no writeback) on cancel */
  69 #define LDLM_FL_MATCH_DATA     0x020000 /* see ldlm_lock_match */
  70
  71 /* These are flags that are mapped into the flags and ASTs of blocking locks */
  72 #define LDLM_AST_DISCARD_DATA  0x80000000 /* Add FL_DISCARD to blocking ASTs */
  73
  74 /* Flags sent in AST lock_flags to be mapped into the receiving lock. */
  75 #define LDLM_AST_FLAGS         (LDLM_FL_DISCARD_DATA)
  76
  77 /* The blocking callback is overloaded to perform two functions.  These flags
  78  * indicate which operation should be performed. */
  79 #define LDLM_CB_BLOCKING    1
  80 #define LDLM_CB_CANCELING   2
  81
  82 #define L2B(c) (1 << c)
  83
  84 /* compatibility matrix */
  85 #define LCK_COMPAT_EX  L2B(LCK_NL)
  86 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  87 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  88 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  89 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  90 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  91
  92 static ldlm_mode_t lck_compat_array[] = {
  93         [LCK_EX] LCK_COMPAT_EX,
  94         [LCK_PW] LCK_COMPAT_PW,
  95         [LCK_PR] LCK_COMPAT_PR,
  96         [LCK_CW] LCK_COMPAT_CW,
  97         [LCK_CR] LCK_COMPAT_CR,
  98         [LCK_NL] LCK_COMPAT_NL
  99 };
 100
 101 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
 102 {
 103        if (exist < LCK_EX || exist > LCK_NL)
 104               LBUG();
 105        if (new < LCK_EX || new > LCK_NL)
 106               LBUG();
 107
 108        return (lck_compat_array[exist] & L2B(new));
 109 }
 110
 111 /*
 112  *
 113  * cluster name spaces
 114  *
 115  */
 116
 117 #define DLM_OST_NAMESPACE 1
 118 #define DLM_MDS_NAMESPACE 2
 119
 120 /* XXX
 121    - do we just separate this by security domains and use a prefix for
 122      multiple namespaces in the same domain?
 123    -
 124 */
 125
 126 struct ldlm_namespace {
 127         char                  *ns_name;
 128         __u32                  ns_client; /* is this a client-side lock tree? */
 129         struct list_head      *ns_hash; /* hash table for ns */
 130         __u32                  ns_refcount; /* count of resources in the hash */
 131         struct list_head       ns_root_list; /* all root resources in ns */
 132         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 133         struct list_head       ns_list_chain; /* position in global NS list */
 134         /*
 135         struct proc_dir_entry *ns_proc_dir;
 136         */
 137
 138         struct list_head       ns_unused_list; /* all root resources in ns */
 139         unsigned int           ns_nr_unused;
 140         unsigned int           ns_max_unused;
 141
 142         spinlock_t             ns_counter_lock;
 143         __u64                  ns_locks;
 144         __u64                  ns_resources;
 145 };
 146
 147 /*
 148  *
 149  * Resource hash table
 150  *
 151  */
 152
 153 #define RES_HASH_BITS 10
 154 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 155 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 156
 157 struct ldlm_lock;
 158
 159 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 160                                       struct ldlm_lock_desc *new, void *data,
 161                                       int flag);
 162 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags,
 163                                         void *data);
 164
 165 struct ldlm_lock {
 166         struct portals_handle l_handle; // must be first in the structure
 167         atomic_t              l_refc;
 168         struct ldlm_resource *l_resource;
 169         struct ldlm_lock     *l_parent;
 170         struct list_head      l_children;
 171         struct list_head      l_childof;
 172         struct list_head      l_lru;
 173         struct list_head      l_res_link; // position in one of three res lists
 174         struct list_head      l_export_chain; // per-export chain of locks
 175         struct list_head      l_pending_chain; // locks with callbacks pending
 176         unsigned long         l_callback_timeout;
 177
 178         ldlm_mode_t           l_req_mode;
 179         ldlm_mode_t           l_granted_mode;
 180
 181         ldlm_completion_callback l_completion_ast;
 182         ldlm_blocking_callback   l_blocking_ast;
 183
 184         struct obd_export    *l_export;
 185         struct lustre_handle *l_connh;
 186         __u32                 l_flags;
 187         struct lustre_handle  l_remote_handle;
 188         void                 *l_data;
 189         struct ldlm_extent    l_extent;
 190         __u32                 l_version[RES_VERSION_SIZE];
 191
 192         __u32                 l_readers;
 193         __u32                 l_writers;
 194         __u8                  l_destroyed;
 195
 196         /* If the lock is granted, a process sleeps on this waitq to learn when
 197          * it's no longer in use.  If the lock is not granted, a process sleeps
 198          * on this waitq to learn when it becomes granted. */
 199         wait_queue_head_t     l_waitq;
 200         struct timeval        l_enqueued_time;
 201 };
 202
 203 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 204 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
 205                                void *req_cookie, ldlm_mode_t mode, int flags,
 206                                void *data);
 207
 208 #define LDLM_PLAIN       10
 209 #define LDLM_EXTENT      11
 210
 211 #define LDLM_MIN_TYPE 10
 212 #define LDLM_MAX_TYPE 11
 213
 214 struct ldlm_resource {
 215         struct ldlm_namespace *lr_namespace;
 216         struct list_head       lr_hash;
 217         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 218         struct list_head       lr_children; /* list head for child resources */
 219         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 220                                              * part of lr_children if child */
 221
 222         struct list_head       lr_granted;
 223         struct list_head       lr_converting;
 224         struct list_head       lr_waiting;
 225         ldlm_mode_t            lr_most_restr;
 226         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 227         struct ldlm_resource  *lr_root;
 228         struct ldlm_res_id     lr_name;
 229         __u32                  lr_version[RES_VERSION_SIZE];
 230         atomic_t               lr_refcount;
 231
 232         /* lr_tmp holds a list head temporarily, during the building of a work
 233          * queue.  see ldlm_add_ast_work_item and ldlm_run_ast_work */
 234         void                  *lr_tmp;
 235 };
 236
 237 struct ldlm_ast_work {
 238         struct ldlm_lock *w_lock;
 239         int               w_blocking;
 240         struct ldlm_lock_desc w_desc;
 241         struct list_head   w_list;
 242         int w_flags;
 243         void *w_data;
 244         int w_datalen;
 245 };
 246
 247 extern struct obd_ops ldlm_obd_ops;
 248
 249 extern char *ldlm_lockname[];
 250 extern char *ldlm_typename[];
 251 extern char *ldlm_it2str(int it);
 252
 253 #define __LDLM_DEBUG(level, lock, format, a...)                               \
 254 do {                                                                          \
 255         if (lock->l_resource == NULL) {                                       \
 256                 CDEBUG(level, "### " format                                   \
 257                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "\
 258                        "res: \?\? rrc=\?\? type: \?\?\? flags: %x remote: "   \
 259                        LPX64"\n" , ## a, lock, lock->l_handle.h_cookie,       \
 260                        atomic_read(&lock->l_refc),                            \
 261                        lock->l_readers, lock->l_writers,                      \
 262                        ldlm_lockname[lock->l_granted_mode],                   \
 263                        ldlm_lockname[lock->l_req_mode],                       \
 264                        lock->l_flags, lock->l_remote_handle.cookie);          \
 265                 break;                                                        \
 266         }                                                                     \
 267         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 268                 CDEBUG(level, "### " format                                   \
 269                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 270                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64\
 271                        "] flags: %x remote: "LPX64"\n" , ## a,                \
 272                        lock->l_resource->lr_namespace->ns_name, lock,         \
 273                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),   \
 274                        lock->l_readers, lock->l_writers,                      \
 275                        ldlm_lockname[lock->l_granted_mode],                   \
 276                        ldlm_lockname[lock->l_req_mode],                       \
 277                        lock->l_resource->lr_name.name[0],                     \
 278                        lock->l_resource->lr_name.name[1],                     \
 279                        atomic_read(&lock->l_resource->lr_refcount),           \
 280                        ldlm_typename[lock->l_resource->lr_type],              \
 281                        lock->l_extent.start, lock->l_extent.end,              \
 282                        lock->l_flags, lock->l_remote_handle.cookie);          \
 283                 break;                                                        \
 284         }                                                                     \
 285         {                                                                     \
 286                 CDEBUG(level, "### " format                                   \
 287                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 288                        "res: "LPU64"/"LPU64" rrc: %d type: %s flags: %x "     \
 289                        "remote: "LPX64"\n" , ## a,                            \
 290                        lock->l_resource->lr_namespace->ns_name,               \
 291                        lock, lock->l_handle.h_cookie,                         \
 292                        atomic_read (&lock->l_refc),                           \
 293                        lock->l_readers, lock->l_writers,                      \
 294                        ldlm_lockname[lock->l_granted_mode],                   \
 295                        ldlm_lockname[lock->l_req_mode],                       \
 296                        lock->l_resource->lr_name.name[0],                     \
 297                        lock->l_resource->lr_name.name[1],                     \
 298                        atomic_read(&lock->l_resource->lr_refcount),           \
 299                        ldlm_typename[lock->l_resource->lr_type],              \
 300                        lock->l_flags, lock->l_remote_handle.cookie);          \
 301         }                                                                     \
 302 } while (0)
 303
 304 #define LDLM_DEBUG(lock, format, a...) __LDLM_DEBUG(D_DLMTRACE, lock, \
 305                                                     format, ## a)
 306 #define LDLM_ERROR(lock, format, a...) __LDLM_DEBUG(D_ERROR, lock, format, ## a)
 307
 308 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 309         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 310
 311 /*
 312  * Iterators.
 313  */
 314
 315 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 316 #define LDLM_ITER_STOP     2 /* stop iterating */
 317
 318 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 319 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 320
 321 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 322                           void *closure);
 323 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 324                            void *closure);
 325 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 326                                ldlm_res_iterator_t iter, void *closure);
 327
 328 int ldlm_replay_locks(struct obd_import *imp);
 329 void ldlm_change_cbdata(struct ldlm_namespace *, struct ldlm_res_id *,
 330                         ldlm_iterator_t iter, void *data);
 331
 332 /* ldlm_extent.c */
 333 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 334 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock **, void *,
 335                        ldlm_mode_t, int flags, void *);
 336
 337 /* ldlm_lockd.c */
 338 int ldlm_server_blocking_ast(struct ldlm_lock *, struct ldlm_lock_desc *,
 339                              void *data, int flag);
 340 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 341 int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback,
 342                         ldlm_blocking_callback);
 343 int ldlm_handle_convert(struct ptlrpc_request *req);
 344 int ldlm_handle_cancel(struct ptlrpc_request *req);
 345 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 346
 347 /* ldlm_lock.c */
 348 void ldlm_register_intent(ldlm_res_policy arg);
 349 void ldlm_unregister_intent(void);
 350 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 351 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 352 void ldlm_cancel_callback(struct ldlm_lock *);
 353 int ldlm_lock_set_data(struct lustre_handle *, void *data);
 354 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 355 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *,
 356                                       struct lustre_handle *);
 357
 358 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 359 {
 360         return __ldlm_handle2lock(h, 0);
 361 }
 362
 363 #define LDLM_LOCK_PUT(lock)                     \
 364 do {                                            \
 365         /*LDLM_DEBUG((lock), "put");*/          \
 366         ldlm_lock_put(lock);                    \
 367 } while (0)
 368
 369 #define LDLM_LOCK_GET(lock)                     \
 370 ({                                              \
 371         ldlm_lock_get(lock);                    \
 372         /*LDLM_DEBUG((lock), "get");*/          \
 373         lock;                                   \
 374 })
 375
 376 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 377 void ldlm_lock_put(struct ldlm_lock *lock);
 378 void ldlm_lock_destroy(struct ldlm_lock *lock);
 379 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 380 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 381 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 382 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 383 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 384 int ldlm_lock_match(struct ldlm_namespace *ns, int flags, struct ldlm_res_id *,
 385                     __u32 type, void *cookie, int cookielen, ldlm_mode_t mode,
 386                     void *data, struct lustre_handle *);
 387 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 388                                         int *flags);
 389 void ldlm_lock_cancel(struct ldlm_lock *lock);
 390 void ldlm_cancel_locks_for_export(struct obd_export *export);
 391 int ldlm_run_ast_work(struct list_head *rpc_list);
 392 void ldlm_reprocess_all(struct ldlm_resource *res);
 393 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 394 void ldlm_lock_dump(int level, struct ldlm_lock *lock, int pos);
 395 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 396
 397 /* ldlm_test.c */
 398 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 399 int ldlm_regression_start(struct obd_device *obddev,
 400                           struct lustre_handle *connh,
 401                           unsigned int threads, unsigned int max_locks_in,
 402                           unsigned int num_resources_in,
 403                           unsigned int num_extents_in);
 404 int ldlm_regression_stop(void);
 405
 406
 407 /* resource.c */
 408 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 409 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 410 int ldlm_namespace_free(struct ldlm_namespace *ns);
 411 int ldlm_proc_setup(struct obd_device *obd);
 412 void ldlm_proc_cleanup(struct obd_device *obd);
 413
 414 /* resource.c - internal */
 415 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 416                                         struct ldlm_resource *parent,
 417                                         struct ldlm_res_id, __u32 type,
 418                                         int create);
 419 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 420 int ldlm_resource_putref(struct ldlm_resource *res);
 421 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 422                             struct ldlm_lock *lock);
 423 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 424 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 425 void ldlm_dump_all_namespaces(void);
 426 void ldlm_namespace_dump(struct ldlm_namespace *);
 427 void ldlm_resource_dump(struct ldlm_resource *);
 428 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 429                               struct ldlm_res_id);
 430
 431 /* ldlm_request.c */
 432 int ldlm_expired_completion_wait(void *data);
 433 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 434 int ldlm_cli_enqueue(struct lustre_handle *conn,
 435                      struct ptlrpc_request *req,
 436                      struct ldlm_namespace *ns,
 437                      struct lustre_handle *parent_lock_handle,
 438                      struct ldlm_res_id,
 439                      __u32 type,
 440                      void *cookie, int cookielen,
 441                      ldlm_mode_t mode,
 442                      int *flags,
 443                      ldlm_completion_callback completion,
 444                      ldlm_blocking_callback callback,
 445                      void *data,
 446                      struct lustre_handle *lockh);
 447 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 448                     void *data, __u32 data_len);
 449 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 450 int ldlm_cli_cancel(struct lustre_handle *lockh);
 451 int ldlm_cli_cancel_unused(struct ldlm_namespace *, struct ldlm_res_id *,
 452                            int flags, void *opaque);
 453
 454 /* mds/handler.c */
 455 /* This has to be here because recurisve inclusion sucks. */
 456 int intent_disposition(struct ldlm_reply *rep, int flag);
 457 void intent_set_disposition(struct ldlm_reply *rep, int flag);
 458 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 459                      void *data, int flag);
 460
 461
 462 /* ioctls for trying requests */
 463 #define IOC_LDLM_TYPE                   'f'
 464 #define IOC_LDLM_MIN_NR                 40
 465
 466 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 467 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 468 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 469 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 470 #define IOC_LDLM_MAX_NR                 43
 471
 472 #endif